Skip to content
Hangfire storage adapter for MySql / fork of not maintained Hangfire.MySqlStorage
C# F# TSQL Batchfile
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
src tweaks Jun 18, 2019
.gitignore project imported May 29, 2019
CHANGES.md
Common.targets SQL typo Jun 2, 2019
LICENSE empty project May 29, 2019
README.md mysql transactions Jun 1, 2019
build.fsx deadlocks treatment May 30, 2019
build.fsx.lock deadlocks treatment May 30, 2019
build.imports.fsx empty project May 29, 2019
build.tools.fsx deadlocks treatment May 30, 2019
fake.cmd empty project May 29, 2019
paket.cmd empty project May 29, 2019
paket.dependencies xa recover Jun 1, 2019
paket.lock xarecover -> net462 Jun 3, 2019
settings.cfg build config May 30, 2019

README.md

Hangfire.Storage.MySql

NuGet Stats

This is MySQL storage adapter for Hangfire.

It is based on Hangfire.MySqlStorage and Hangfire.MySql.Core.

Story

It is relatively complicated.

As far as I understand Hangfire.MySqlStorage was first (and is listed on Hangfire website as 3rd party driver). It was not released for .NET Core (.NET Standard) for a very long time though so unofficial fork has been created, Hangfire.MySql.Core, which main purpose was: .NET Core. It also included few fixes but nothing revolutionary. After a while original Hangfire.MySqlStorage caught up with .NET Core and some fixes (nothing revolutionary either).

Unfortunately, there are still more serious bugs which reveal themselves under load.

If you see exceptions saying Too many connections, Deadlock found when trying to get lock; try restarting transaction, Transaction branch was rolled back: deadlock was detected, or something like Timeout expired you are probably having concurrency problems related to deadlocks.

You can find some signs that authors tried to address those issues, like some SET SESSION TRANSACTION ISOLATION LEVEL READ UNCOMMITTED here and there but this is not solving problems in high-load situations.

Identified concurrency problems

There are actually 3 scenarios I personally encountered:

  • Single statement dead-lock with exception (MySQL error #1213)
  • Transaction dead-lock with exception (MySQL error #1614)
  • Transaction (only?) (true) dead-lock freezing whole operation

As first two are relatively benign (we could just catch exception and retry) the third one is deadly. This is just a speculation, but it seems like it locks tables permanently, so even killing application is not removing locks from tables. This is weird and I don't understand it completely, so maybe I'm wrong about the diagnosis (but not about symptoms).

Deadlocks caused by poor performance

They are not really caused by poor performance, but poor performance increases window of opportunity, so deadlock caused by other bugs have greater chance to manifest themselves.

  • Missing index for Job.ExpireAt and Set.ExpireAt (but also other tables) which kills ExpirationManager (note: without index on ExpireAt delete operation leads to full table scan and locks all rows)
  • Missing index for Set.Score which kills MySqlStorageConnection.GetFirstByLowestScoreFromSet (note: as name suggests it is about lowest score, doing full table scan on with >1m rows every second has major impact)
  • Missing index for Job.FetchToken which slows down MySqlJobQueue.Dequeue (note: in my case list of jobs to be done NOW was quite long so this index was important)

NOTE: This library modifies database slightly (adds some indexes), so after running it for the first time you have slightly different database than original, for better and for worse.

Deadlocks caused by no preemptive locking

MySqlWriteOnlyTransaction seems to have an idea of locking resources modified by transactions (methods like: AcquireJobLock) but actual implementation is empty, so when transaction starts nothing is locked, so MySQL locks whatever MySQL wants to lock, in random order leading to complete standstill and leaking connections.

I hoped for catching MySQL error #1614 and retrying transaction but excepion was not always thrown, sometimes it was just freezing.

So preemptive locking was the safest but.

The problem with preemptive locking is, that if you decide to use them, you have to use them everywhere. So, all inserts, updates, and deletes are now using preemptive locks on appropriate resources.

Deadlocks caused by distributed transactions

Implementation of TransactionScope for MySqlConnector uses MySQL distributed transactions (XA). It is a slightly problematic as such transaction survive crashes, shutdowns, etc. As Hangfire never reconnects to transaction, those hanging transaction pile up and lock tables. It took me few days to find what is the problem. See XaRecover console app in this repository which can be used to lift all distributed transactions.

Usage

As in original Hangfire.MySqlStorage:

var options =
    new MySqlStorageOptions {
        TransactionIsolationLevel = IsolationLevel.ReadCommitted,
        QueuePollInterval = TimeSpan.FromSeconds(15),
        JobExpirationCheckInterval = TimeSpan.FromHours(1),
        CountersAggregateInterval = TimeSpan.FromMinutes(5),
        PrepareSchemaIfNecessary = true,
        DashboardJobListLimit = 50000,
        TransactionTimeout = TimeSpan.FromMinutes(1),
        TablesPrefix = "Hangfire"
    };
var storage = new MySqlStorage(connectionString, options);

In ASP.NET Core you register hangfire in services:

services.AddHangfire(config => config.UseStorage(storage)));

but for example, in command line application, you just create BackgroundJobServer youself:

var server = new BackgroundJobServer(storage);

there is also global configuration:

GlobalConfiguration.Configuration.UseStorage(storage);

There rest is in (capable) hands of Hangfire.

Next

I tried to keep modifications as minimal as possible to make merging back to Hangfire.MySqlStorage as easy as possible, and once changes are merged I could go back to gardening. But... It seems like all issues are not answered, several pull requests are not merged, and nobody knows if they ever will be. The code is full of deprecation warnings and ReSharper issues so my fixes are just a beginning, and there is a lot of work to be done, to polish it.

If I decide to keep developing it, I guess, it might be better to fork it from official Microsoft SQL Server implementation.

Build

paket install
fake build
You can’t perform that action at this time.